Precision Grounding: Augmenting Large Language Models with Evidence-Based Databases for Trustworthy Genetic Variant...

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Accurate genetic variant interpretation is crucial for advancing precision medicine.
• Large language models (LLMs) show potential for summarizing genetic data but are susceptible to hallucinations.
• Existing retrieval-augmented generation (RAG) methods often rely on less specific document embeddings.

Purpose of the Study:

• To introduce a novel
• precision grounding
• approach to enhance LLM accuracy for genetic variant summarization.

Main Methods:

• Developed CATT, an open-source tool integrating ClinGen, ClinVar, and GenCC data for variant-specific evidence retrieval.
• Utilized a domain-specific query tool to access evidence-based databases via unique identifiers, unlike traditional RAG.
• Compared precision grounding with web-search grounding using 50 expert-selected genetic variants and GPT-4o.

Main Results:

• Precision grounding significantly outperformed web-search grounding, achieving higher accuracy (4.76) and completeness (4.94) scores.
• Error analysis confirmed a reduction in clinically significant hallucinations, including incorrect pathogenicity classifications.
• The approach effectively grounded LLM outputs with curated, variant-specific evidence.

Conclusions:

• Precision grounding represents a significant advancement for accurate genetic variant summarization using LLMs.
• The CATT tool provides a practical solution for integrating domain-specific knowledge to mitigate LLM hallucinations.
• This method holds promise for improving the reliability of AI-driven tools in precision medicine.